ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

graph : normalize Q, K, V shapes + sync cross attention (#12449)

Browse source 
* graph : normalize Q, K, V shapes and add comments

ggml-ci

* context : synchronize before getting cross attention data

* model : fix command-r attention norm check
This commit is contained in:
Georgi Gerganov 2025-03-18 21:35:19 +02:00 • committed by GitHub
parent bb115d2bf7
commit 75422e8bc4
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 433 additions and 277 deletions
Download patch file Download diff file Expand all files Collapse all files

24
src/llama-graph.h
View file

@ -487,9 +487,9 @@ struct llm_graph_context {
ggml_tensor * build_attn_mha(
ggml_cgraph * gf,
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * q, // [n_embd_head_q, n_tokens, n_head_q]
ggml_tensor * k, // [n_embd_head_k, n_tokens, n_head_k]
ggml_tensor * v, // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
ggml_tensor * kq_b,
ggml_tensor * kq_mask,
bool v_trans,
@ -502,9 +502,9 @@ struct llm_graph_context {
ggml_cgraph * gf,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
float kq_scale,
int il) const;
@ -516,9 +516,9 @@ struct llm_graph_context {
ggml_cgraph * gf,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
float kq_scale,
int il) const;
@ -530,9 +530,9 @@ struct llm_graph_context {
ggml_cgraph * gf,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
float kq_scale,
int il) const;